JPS60254100A - Voice recognition system - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION (Technical Field) The present invention relates to a speech recognition system, and particularly to speech segment detection for detecting speech segments with high accuracy.

(Background Art) A block diagram of a conventional speech recognition device is shown in FIG. 1st
In the figure, ■ is a signal input terminal, 2 is a frequency analysis section, and 3 is a signal input terminal.
4 is a voice capture control section, 4 is a capture start signal, 5 is a voice section detection section, 6 is a capture end signal, 7 is start/end information, 8 is a recognition section, and 9 is an output terminal. 9. Each part will be explained below.

The frequency analysis section 2 is configured as shown in FIG.
The human voice signal 11 is amplified to an appropriate level by a preamplifier 12, and is then amplified by a group of N band filters i3 and a group of full-wave rectifiers, which are equally spaced on a logarithmic scale from approximately 201 JHz7 to 6000 Hz. 14, and a group of low-pass filters 15, and furthermore, the multiplexer 16 is sequentially switched at every predetermined time period (hereinafter referred to as a sample period), and then quantized by the AD converter 17. The analysis result 18 is output.

After receiving the capture start signal 4, the voice capture control unit 3 sends the analysis result 18 of the frequency analysis unit 2 to the voice section detection unit 5 and the recognition unit 8 for a certain period of time, and then the voice input is completed exactly. To determine when the audio input has ended, for example, the average value of N data for each sample period (hereinafter 7, V-mye-wah,!: Record-
r) After a certain number of sounds whose power exceeds a preset threshold, it is determined that voice input has ended when a certain number of sounds that do not exceed the threshold continue. There is a way.

A block diagram of the voice section detection section 5 is shown in FIG.

In FIG. 3, 18 is an analysis result, 21 is a parameter calculation unit, 6 is a capture end signal, 22 is a blocking unit, 23 is a voice section determination unit, and 7 is configured as start and end information, which will be explained in detail below. do.

The parameter calculation unit 21 is a part that calculates parameters defined by the expression (]) used for voice section detection from the analysis result 18.

P-a, ·
; Average value of the third analysis results Also, the gradient a, that is, the slope of the least squares approximation straight line, is the average value of the third N analysis results. (numbers assigned in descending order of frequency of the filter group), a is determined by equation (2).

can be replaced, and equation (2) is transformed into equation (3).

I can do it.

Also, n is obtained by dividing ΣX 1J by N! =1 can be done. FIG. 4 is a block diagram for calculating Pl, and will be explained below according to the diagram.

Third N analysis results x r , + ('-1,2
r...N) are output in order, the result of the adder 101 is output to the multiplier 103 and the divider 106.

Yes, the complementer 104 calculates the i=+1J value of -cl·ΣX and inputs it to one side of the adder 1.05.
Counter 107OIj that works in synchronization with the data output of 1J
j force, Xlj and) NR' −x l J multiplier 10
8, an adder 109 connected to the output of the multiplier 108, and an adder 110 connected thereto.
G, te1, 1'°゛, J to 7', member C) can be done. Output of register 10, δ+ −x r j
The multiplier 111 is connected to one input of the multiplier 111, and the other input of the multiplier 111 is set to N, and the multiplier 111 calculates and adds The signal is input to the other side of the device 105. The adder 105 calculates the spread torque slope of the sample data &4. 1.1 The result becomes one input of the multiplier 113. Also, the divider 1067 is
The result becomes the other input of the multiplier 113, and the multiplier 113 can multiply P, (-a-x,). By performing the above calculations for each sampling period, all values of P at each sampling time can be calculated.

The blocking unit 22 uses the result PJ of the parameter calculation unit 21.
The block diagram in FIG. 5 shows the part where the audio is processed until the capture end signal 6 is detected, and after the capture end signal 6 is detected, the audio is professionally converted (detection of chunks of parts that seem to be audio). A diagram is shown and the explanation will be given according to FIG.

, for each sampling period of the e-y meter calculation unit 21 (D P
j -: 1, which are sequentially stored in the p-crameter memo +J 200, are read in order, converted into absolute values by the absolute value circuit 201, and input IP, l to one side of the comparator 202. The other input of the comparator 202 is set to an IP- threshold PTH. In the comparator 202, the IP
When 71≧PTH, α output, IPj I <PT
When the signal is H, a significant signal is output to each β output.

The counter 203 counts up when IPjl≧PTH, and is cleared when IPjl<P, H, and counts the continuous amount that satisfies IP−≧PTI. Also, the output of the counter 203 is as follows: always register 20
It is set to 4. The value set in the register 204 (IP, consecutive number where l≧PTH) is set in the comparator 2
05, and K is set in the other input of the comparator 205, and when IP-≧PTl (a continuous quantity (hereinafter referred to as block Chokichi)) or more, the output C of the comparator 205 is significant. A signal is output.

The block length is K (a natural number of K≧2) or more (when outputting C signal), and the β output of the comparator 202 (1 pJl < PT
The timing at which H) appears is captured by the AND circuit 206. The counter 207 counts the amount of P read out at the output size from the output of the AND circuit 206. By subtracting the result (block length) of the register 204 from the output of the counter 7 by the H reducer 208, the block length is calculated. You can measure the distance (time) between The subtractor 21.0 counts the counter 209 in synchronization with the reading of P.
By subtracting the output (block length) of the register 204 from the result of step 09, the beginning of the block can be determined.

IPjl≧PTl by adder 211 and register 212
The accumulation of the □ part indicates the size of the pro.
When the signal of the AND circuit 206 is detected, it is set in the register 213 and at the same time, the signal of the AND circuit 206 is set.
3 (hereinafter referred to as block amount), the output of the subtracter 210 (block head information), the output of the register 204 (block length), and the iJJ force (interblock distance) of the subtractor 208 are stored in the professional table 214. register. In this way, all the imported Japanese can be divided into blocks.

The speech section determination section 23 determines the speech section from the block table 214 obtained by the blocking section 22 in the following manner. In other words, the block with the maximum block amount is detected, and with respect to the blocks before and after the block as the center of the voice section, if the distance between the blocks is less than a certain value, the block is also included in the children's voice section.
The voice section was being determined.

The recognition unit 8 sends a capture start signal to the voice capture control unit 3, stores the analysis results from the voice capture control unit 3, and further receives start and end information 7 from the voice section detection unit 5. The similarity calculation is performed between the standard pattern prepared in advance and the contents of which are known, and the standard pattern with the highest degree of similarity is selected.
It determines that the audio with the same content as the turn is input, and outputs the result.

However, in the voice section detection in the conventional technique described above, (1) the spectral slope aj changes depending on the strength of the input voice, so that the parameter Pj is unstable.

(2) The spectral slope aj changes depending on the phoneme and the speaker, as well as the characteristics of the microphone, and takes a value close to 0 even in the voice part, and as a result, P also takes a value close to O, which makes it difficult for professionals and listeners to make a mistake.

(3) When the noise is large, it is difficult to distinguish it from the noise (especially consonants).

There was a drawback.

(Problem to be solved by the invention) The purpose of the present invention is to provide a speech recognition method that can improve the recognition rate by eliminating recognition errors. By subtracting turns, this method provides a means for detecting voice sections with higher accuracy and increasing the recognition rate, and will be described in detail below.

(Structure and operation of the invention) FIG. 6 is a block diagram of the present invention, 100 is an input terminal, 200 is a frequency analysis section, 300 is a logarithmic conversion section, 4
00 is a tone conversion section, 500 is a speech interval determination section, a logarithmic conversion section data section 501, a noise pattern detection section 502, a subtraction circuit 503, a multiplication circuit 504, an addition circuit 5
05, division circuit 506, P) parameter memory 507,
Comparator 1 508, FLAG 509, smoothing 1
510, Smoothing 2 511, Blocking 5] 2,
Comparator 2513, block determination 514, voice section determination 5
15, MAXBLK table 516, 600 is a resampling section, 700 is a distance calculation section, 800 is a standard pattern memory, 900 is a determination section, and 1000 is a recognition result output terminal.

In such a configuration, an input audio signal inputted from the input terminal 100 is inputted to the frequency analysis section 200, frequency-analyzed as a regularized signal U(i, j) corresponding to a plurality of frequency bands, and then processed by the logarithmic conversion section. Sent to 300.

The data sent to the logarithmic conversion section 300 becomes spectral information, . . . , word information, etc.; Information will be sent.

The logarithmic conversion unit 300 performs division according to equation (4).

Let u and D be the frequency analysis data.

U(i, D ””1~19 j=1~■O≦u(1,D
≦2047 Let the logarithmic conversion data be ■(i, D.

V(i, j) i'9J=1~■ Here, i indicates the frequency (]ch to 19ch), and J indicates the time (1 frame to (1) frame). Input from the preprocessing section Let the data be u(1,D.U(1,;
) +-1~19 j=1~ω O≦U(11j)≦2
047 Let NB be the number of logarithmic conversion bits. Here N.H.
-4.

U (i, j) > 0 Here, input cover turn POW (J) and input
Equation (5) is the calculation formula for the 10th frame of the turn.

It is defined by equation (6).

k −=’(J-1)/] O+1 However, it is assumed that j=(k-]) 10+1.

The noise level is defined by equation (7).

When the noise level measurement interval is set to 2 and to 2, however, k2 is set to +2.Then, when the cutout slice level L1 is L1 = NLEVEL+LO, POWlo(k3) is larger than Ll for the first time.POW10 Assuming that the frame Jlwoj+=(k31) is 10+1-40, which is 3 to 40 frames backward at the point where (k3 + 1) is larger than Ll, the tentative voice start h1 frame 5TFR1 is converted into STF.
Let it be Rl-Mi'A(J, 1).

Termination detection is 1 (4 or greater than 2+1 and POW
], 0 < k4 ) is smaller than or equal to Ll, the temporary audio end frame EDFR1 is set to EDFR
1=(k4-1)*]O-]+9.

Now, the logarithmic transformation data calculated by the logarithmic transformation unit 300 is
The data V(i, j) is -1 logarithmically transformed data section 501
After being sent to the noise pattern 7 NPAT(i), the noise pattern detection unit 502 divides the noise pattern 7 NPAT(i) in order to calculate the noise pattern 7 NPAT(i). However, when the noise level measurement interval is set to 2 and -2, the values of J2 and J3 are calculated using equation (8).

The formula for calculating the noise pattern NPAT(i) is shown in formula (9).

j=sTFR1 to EDFR1 Next, in the subtraction circuit 503, the multiplication circuit 504, the force ll calculation circuit 505, and the division circuit 506,
In the turn detection section 502, NPAT(i), which is one from equation (9), is used to subtract the noise and turn, and then divide the noise and turn by equation 00).

P(,i)=upper r ((:V(i,,j)−NPAT(
i)/4)2→-9(10)]9□Te・.

2-J"-, (i[J) P(,i) obtained from the equation is stored in the P parameter memory 507, and the comparator] 508 performs a comparison according to the following equation (11).

In the first type, when the slice level L2 is larger than P(j), Fr, Ac(j)-o is set. Also, if L2 is equal to or smaller than P(j), FbAG(
J) = 1. F determined in equation (11)
The value of LAG(J) is stored in FLAG 509 and
Depending on the value of LAG(J), it is sent to smoothing 1510 or smoothing 2511. In smoothing 15''O, perform the operation in the case of FLAG(jl20, FLAG(,i-]) = 0, and FLAG(J+
1) - 0, set FLAG(,r)=o.

Smoothing 25" 1 is FLAG (:r)
= 1, and when FLAG (j-] ) = 1 and FLAG (j+] ) -1, FLA
Let G(j) be −1.

Next, in blocking 512, FLAG(J) = 1
is continuous for 4 frames or more, and pow 1 of that section
(z) −yoΣ:P(j) is p□w 1 (/−)≧L
3, i.e. POW 1. A block is represented when (/i) is thicker than or equal to slice level L3.

The first frame f of block l is 5(t), and the last frame of block l is E(,4).
What should I do? Subtracted the noise and P turn of block l.
The added value of the power P (j) is determined by equation (12).

The number of frames in block l is determined by equation (13).

FR](t)2E(71-5(t)+1 ・ ・ (
13) Sunda, former pro, the distance with Ri (1-1) is the (14th)
) is determined by the formula.

=2(t)=s(t)-E(1-1,)...(14)
Here, tl is the audio beginning block, and 12 is the audio final block. In the comparator 2 513, as long as the audio beginning process, tl, satisfies the condition of equation (I5), t = t1-1. do.

FR2(4)≦iV[IN(POWI(4))/SC1
+SC2, 5C3) - (15) For the final audio pro, ri12, set as J-2-t2+1 as long as the condition of equation (1G) is satisfied.

FR(22+1.)≦MIN(POWI(tz+1)/
SC]+SC2,5C3)-f16) Here SCI~S
C3 is a constant and SC is 16, SC228゜5C3
-30.

Based on the above formula, it is determined whether or not the blocks before and after the largest block are to be included as part of the voice section, and are adopted as the voice section.

The values of the audio optical [phase] block L1 and the audio final block t2, which are candidates for the audio area 1141, determined in this manner, are sent to the block determination 514.

Next, the table MAXB LK table 1118 for the maximum number of blocks of recognition words used for speech section determination 4]7 will be explained.

An example of the maximum number of blocks MAXBLKO is shown in FIG.

The left side shows power lever 'IJ (16 words), the right side tqt:,
It shows the maximum number of blocks for each force lever determined in advance from the vocalization data. MAXBL is the largest among these recognition word sets.
Give K. For example, if the recognized word contains "Moichido J", it will be 1VIAXB LK23.

In the voice section determining unit 515, when BLKS≦MAXBLK, that is, when the number of blocks BLKS is smaller than or equal to the maximum number of blocks MAXBLK, all blocks are determined to be voice sections. On the contrary, BLKS)
When MAXBLK is set, that is, the number of BLK is maximum.

If the number of blocks MAXBLK is 1: If the number of blocks is also large, for example (in Figure 7, the number of blocks BLKS = :3), if there are two major professionals, and the number of blocks MAXBLK = 2, the introduction of ■ or ■ is possible, and ■ After determining the power PP (t) of each of the combinations of blocks (2) and (2), the PPs are compared, and the combination of blocks for which the professional power PP (t) is maximized is determined as the voice section. Block power PP(t)
is determined by equation (17).

t=1~BLKS-MAXBLK+1 S (/l-+) determined from equation (I7) is the audio beginning block, rear edge, E (22) is the audio final block,
The voice start frame 5TFR is -5TFR,=S(L,), and the voice end frame EDFR is EDFR2E(, -52). The input/turn frame number IFR is expressed by the following equation (18).

TFR2EDFR-5TFR+1 (18) The determination of the end of the process is made when the final audio block t2 satisfies all the conditions of the following formula.

That is, L 1 is 1<4, 4'-1-1, k4+
2, k4-1-3.

If it is thick or equal to any of k 4 +4 +, the process ends.

-1 If the condition of formula (I9) is not satisfied,
Abort recognition pow], o(k4)≦L1 In other words, set the value of 4 to the next value where Ll is greater or equal.

Voice section 5TFR and E[)FR determined in this way
is W(i,j) sent from the spectrum conversion section 400
At the same time, it is sent to the re-sampling unit 500. The resampling unit 500 normalizes the time axis of the audio. The method of normalizing the time axis is a conventionally known technique, and is a linear matching method (a method in which the speech interval is divided into a fixed number determined by the conditions of the recognition device at equal intervals in time and resampled). It is.

Then, the distance calculation section 700 performs a distance calculation with the output of the standard baton memory 800 created in the same manner, and sends the result to the determination section 900.

The determination unit 900 compares the distance value with the total distance, and outputs the category name with the smallest total distance as the recognition result from the recognition result output Mizuko 1000.

As mentioned above, in the present invention, by subtracting noise and patterns from the speech pattern when detecting a speech section, the speech section can be detected more accurately and the recognition rate can be increased.

(The following is a margin) (Effects of the Invention) The present invention subtracts information on the noise pattern of the voice from the voice pattern information when detecting the voice zone, thereby making it possible to detect the voice zone more accurately and improve voice recognition. It is effective in improving the recognition performance of the device.

[Brief explanation of drawings]

Fig. 1 is a block diagram of a conventional speech recognition device, Fig. 2 is a detailed block diagram of the frequency analysis section shown in Fig. 1, Fig. 3 is a block diagram of the speech interval detection section shown in Fig. Figure 3 is a detailed block diagram of the parameter calculation section, Figure 5 is a detailed diagram of the programmerization unit in Figure 3, Figure 6 is a detailed diagram of the speech recognition device of the present invention, and Figure 7 is FIG. 8 is a diagram showing the combination of blocks in the audio section, and FIG. 8 is a diagram showing the maximum number of audio blocks. 1. Input terminal, 2. Frequency analysis section, 3. Audio capture control section, 4. Capture start signal, 5. Voice section detection section,
6... Capturing end signal, 7... Starting end/terminating end information, 8...
... Recognition unit, 9... Output terminal, 11. Input audio signal,
12... Preamplifier, 13... Band pass filter group, ■4... Full wave rectifier group, 15... Low pass filter group,
16... Multiplexer, 17... AD converter, 18...
...Analysis results, 21...Parameter calculation section, 22...
Blocking unit, 23・Voice section determining unit, 101, 1.0
5゜109...adder, 102,110...register, 103.108,111,113...multiplier, 10
4 Complementer '1 106.112...Divider, 107
...Counter, 200・P parameter calculation section J, 2
01・Absolute value circuit, 202, 205・Comparator, 203,
207,209・Counter, 204. ．． 212, 21
3. Register, 206...AND circuit, 208, 210
Subtractor, 211 - Adder, 214 - Block table, 100 - Input terminal, 200... Frequency analysis section, 3'
OO: Logarithmic conversion unit, 400: Spectrum conversion unit, 500: Voice interval determination unit, 501: Logarithmic conversion unit, 502: Noise/tan detection unit, 503: Subtraction circuit,
504・Multiplication circuit, 505・Addition circuit, 506...
Division circuit, 507・P parameter memory, 508...
・Comparator 1.509...FLAG,, 510...Smoothing 1.511...Smoothing 2.512...
・Blocking, 513... Comparator 2', ) 12 ,,
514...Block determination, 515...Speech section determination, 516...MAXBLK,, 600...Re-sampling section, 700-distance calculation section, SOO...Standard...
900... Judgment unit, 1000... Recognition result output terminal. Patent applicant Oki Electric Industry Co., Ltd. Patent application agent Megumi Yamamoto - Figure 1 Figure 2 Figure 3

Claims (1)

[Claims] Analyze the frequency of the input audio signal, logarithmically transform the result,
The analyzed spectral characteristics are normalized and resampled to a certain data length to create the input audio pattern.
In a speech recognition method that calculates the distance between the pattern and the standard pattern and recognizes and determines the recognition category with the minimum distance; means for storing logarithmic transformation data; means for calculating the power obtained by subtracting the noise pattern from the voice/guitar; means for setting a voice section flag from the calculated power information and performing smoothing;
A speech recognition method comprising: means for determining a speech block candidate based on a speech section flag obtained by the processing; and means for determining a speech section using the speech block candidate while referring to a maximum block (MAXBLK) table. .